Stabilizer
Name is misleading. This is NOT a LoRA that can magically fix your overfitted merged base model.
This is a finetuned model from pretrained model. But trained as a LoRA.
11k images. No bias, no default style.
Zero smooth plastic glossy AI image in dataset. Glossy Al images are polluting the world, but not on my watch, I handpicked every single image.
Natural language captions from Gemini, rather than tags in random order with high FPR.
Standard noise scheduler. Inpainting/img2img friendly.
Improvements:
You can get the style exactly as it should be (as long as the model knows).
Comparisons with pretrained base model:
https://civarchive.com/images/84145167 (general styles, you get the style exactly as it should be)
https://civarchive.com/images/84256995 (artist styles, no style shifting because this model does not have bias)
See more xy plots in cover images.
Why LoRA?
This LoRA is a DoRA (from Nvidia), which is more efficient than traditional LoRA.
Enough for thousands of training images.
What? You prefer to download and store a 7GiB checkpoint than a 80Mib LoRA?
Share merges using this model is prohibited. FYI, there are hidden trigger words to print invisible watermark. I coded the watermark and detector myself. I don't want to use it, but I can.
This model only published on Civitiai and TensorArt. If you see "me" and this sentence in other platforms, all those are fake and the platform you are using is a pirate platform.
How to use
It is highly recommended that you use pretrained base model.
And load this LoRA with strength 1.
Versions:
cknb (ChenkinNoob-XL).
nbvp10 (NoobAI v-pred v1.0). FYI: you don't need CFG hacks (RescaleCFG etc.).
nbep10 (NoobAI eps v1.0).
illus01 (Illustrious v0.1).
Load this LoRA first
This LoRA uses a new arch called DoRA from Nvidia, more efficient than traditional LoRA. But the patch weight is dynamically calculated based on the currently loaded base model weights (which will be changed when you loading other LoRAs). To avoid unexpected changes, load this LoRA first.
Specify styles in prompt
This model does not have an strong default style and is very creative. You must specify the style you want in the prompt.
If you want to use it on finetuned/merged base models:
I personally disagree this. This is not the model's original intention. But this is a LoRA after all.
Be aware:
This LoRA can't remove glossy shiny plastic AI style. Although the dataset is "AI image free". It can't make AI image polluted 1girl overfitted 50 versions of Nova furry 3D anime WAI or whatever look better. What the model learned is learned. Use a pretrained base model if you want to get rid of overfitted AI style.
What is "overfitted AI style"? This is what Craft Lawrence (from spice and wolf) should be, if you've seen the anime: img. This is what those AI style polluted 1girl overfitted model generated: img
Some base models already merged this model. If you got deformed images with this LoRA even at low strength (e.g. <0.5). Your base model has already merged this LoRA (and you merged it twice). And the model weights got multiplied (because how the DoRA works) and collapsed.
Beware of fake base model creators, aka. thieves. Some "creators" never do the training, they only grab other people's models, merge them, wipe all metadata and credits, and sell it as their own base model.
Update log
Moved here
Description
FAQ
Comments (31)
Love this LoRA. Thank you for continuing to improve it!
Friendly reminder: WAI models are also NoobAI based. And v14/15 is not based on illustrious v1.
v15 cos similarity against NoobAI eps v1.0 is 0.995. and against illustrious v1 is 0.94.
V14 is the same.
As a ref: similarity between NoobAI and illustrious v0.1 is ~0.94.
I assume it merged NoobAI eps v1 and v0.5. Because model weights are somewhere between those two checkpoint, closer to eps v1
One thing for sure, it's not illustrious v1, not even close to. illustrious v1 is on the other side of the Earth.
Why the author lies and refuses to give NoobAI credit?
don't ask me. I don't know.
¯\_( ͠° ͟ʖ °͠ )_/¯
What strenght would you suggest to be used with WAI 15?, ive been doing some test today with different weights, first trying the suggested by the page of "Strenght:1" but it cooked the image. so im trying with 0.3 and 0.2. But i would like to know your thoughts if possible, please :')
@Razuryam it depends what you want.
if you want better and more logical background. I think 0.3 is good.
If you want natural lighting, details, textures. It would never work on wai because wai models are trained on AI images and have strong AI style.
@reakaakasky Oooh i see, thanks for letting me know. I dont understand much of the topic of how the models are trained or whats the difference between them. Most of the time i just download and start doing images hehe... I know, i know i do the Unga boonga way haha. Ty!
@Razuryam I'd just avoid Wai15 entirely, unless you enjoy inpainting fingers. Subjectively, I feel like everything past v9 has been a downgrade, but v15 is especially bad at details.
@EbenezerDanglewood oh thanks for the advice, and sorry for the SUUUUPER Late reply. I find myself very absend lately with the site lmao.
we need you to update your tensorArt library! :)
all my problems disappeared when i stopped using this lora
233333333
that bad eh
Does this not work with Janku v5 noobai illustrious? It's just making my images look terrible, even on lower weight I can't say it's better.
janku model has strong AI style, it will suppress other LoRAs and make them almost have no effect.
but if you said things became "terrible", like color blobs, are you using the right LoRA version?
I think you used wrong version. It works perfectly, unless you used v-pred version.
@reakaakasky I used "noobai_vp10_stabilizer_v0.271_fp16" and it looks like that first picture you posted yes. I will try the other one, thank you, btw you said the pred one was better, how come that one doesn't work and the other one does? Why does it give such a weird effect? Is that only with Janku?
@Doozyy I mean v-pred models are better, not just this LoRA.
Why does it give such a weird effect? Is that only with Janku?
v-pred and eps are two different types of model. they can't be mixed.
Finally tried one of these on Noob Vpred, love the result.
enhancer for lunima 2 https://civitai.com/models/1974130
You are creating finetune of Lumina 2.0?
Its nice, but how about heavy censorship? Gemma 2 is censored to hell and back?
@Mescalamba I don't do nsfw, but I did simple test on Netayume. I think it's ok.
Whether a TE model is censored or not, does not matter. What does matter is that the TE must know those concept.
In fact, censorship maybe good. Because it means the TE must know those concept very well, so it can refuse it later(as a LLM). But we use it as a encoder, so it means it may be better
@reakaakasky Gemma 2 2B is decoder only model with fused "CLIP" equiv to top layer of that LLM. But since it does take system prompt, it could be soft-hacked via jailbreak (can try later).
There isnt much encoder-decoder models used for instruction, basically only T5.
@Mescalamba text encoder only needs to distinguish those nsfw concept, and the diffusion model generates the concept.
naming is hard. most of image model papers usually just call it text encoder, for the perspective of the diffusion model. Even if it is a decoder-only transformer...
The "illustrious" ones don't seem to work on illustrious.
I don't know how to read. Can someone explain what is the purpose of a stabilizer?
Actually crazy useful for how I do inpainting, already reducing headaches on the first pic I used it for
Pure sorcery alongside the v-pred fix from the control lora collection.
What I'm running into occasionally is this: depending on the prompt, the lora turns the image into a weird, realistic 3d style. specifying the (non-3d) series and artist doesn't seem to work. prompting for 3d and realistic in the negatives also has no effect. The only thing that does is reducing the lora strength, but then I would really appreciate a way to get its full benefits without that quirk.
Is there a style prompt that avoids this issue while preserving the original style of the series/artist?
Oh, and I'm running NoobAI v-pred + v-pred fix and noobAI detailer.
usually if you want to stack multiple LoRAs, you can't go all full strength. There are conflicts when stacking LoRAs.
For example, this LoRA+ v-pred fix might add too much "realistic" effects if their strengths are both too high.
@reakaakasky Yeah, the v-pred fix is at half strength. I've also ditched the detailer because it seemed to be steering towards the realistic thing. For some characters, putting "anime style" or "traditional media" works, but it still commonly struggles to beat the realistic direction. Do you have a prompt recommendation for such cases?
@m4rbleye specific style words, like artist tags.
@m4rbleye https://civitai.com/images/107708598 v-pred quality tags can be pain in ass
@reakaakasky @Volnovik After a deep dive into Danbooru I've found something that works, like, 90% of the time: anime coloring. This prompt seems to simply stick the image to 2d, without having to adhere to a specific artist or style.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.